Page 107 - My FlipBook
P. 107
Brochure 2020
Based on the results from our crash-deterministic SSD, we will further propose a highly e cient
and reliable graph-processing system. Since graph computing entails heavy data trafficking
between memory and storage layers, we will focus on developing techniques to hide or even
eliminate extensive data movements to boost the performance of our graph-processing
system, while being aware of graph access behaviors. Access patterns incurred by graph
computing are mainly in uenced by graph layouts and graph algorithms. Most graph layouts
are constructed by a set of nodes and edges. Graph layouts in the scenarios of social networks
and the World Wide Web usually follow the power law and are termed "scale-free graphs".
For each scale-free graph, some hub nodes have tens or even hundreds the number of edges
than other nodes. The access frequency of each node on performing graph computations is
strongly correlated with the number of edges connected to the node. Thus, hub nodes may
endure tens or even hundred more access events compared with other nodes. Accordingly,
data movements between memory and storage layers can be reduced if we can efficiently
recognize the hub nodes, and prefetch/pin (or lock) the hub nodes in the main memory.
Consequently, we are proposing hub-node pinning strategies to pin recently used hub nodes
in the memory. To avoid wasting the memory resources and prevent high memory access
misses, we will predict and unpin recently unused pinned hub nodes with being aware of the
dynamic behavior of graph access events.
We will also propose a vertical integration solution to integrate the proposed crash-
deterministic SSD and validated SSD management system with dynamic graph applications,
which change nodes and edges through time. That approach of enhancing the performance
of SSDs by simplifying validation complexity and accounting for graph access behaviors will
ensure that graph data can be quickly and correctly recovered when a system encounters
sudden power failures or crashes. At the end of this three-year project, our GraphStor system
will integrate all of these research outputs, representing a highly e cient and reliable graph-
processing system.
105
Based on the results from our crash-deterministic SSD, we will further propose a highly e cient
and reliable graph-processing system. Since graph computing entails heavy data trafficking
between memory and storage layers, we will focus on developing techniques to hide or even
eliminate extensive data movements to boost the performance of our graph-processing
system, while being aware of graph access behaviors. Access patterns incurred by graph
computing are mainly in uenced by graph layouts and graph algorithms. Most graph layouts
are constructed by a set of nodes and edges. Graph layouts in the scenarios of social networks
and the World Wide Web usually follow the power law and are termed "scale-free graphs".
For each scale-free graph, some hub nodes have tens or even hundreds the number of edges
than other nodes. The access frequency of each node on performing graph computations is
strongly correlated with the number of edges connected to the node. Thus, hub nodes may
endure tens or even hundred more access events compared with other nodes. Accordingly,
data movements between memory and storage layers can be reduced if we can efficiently
recognize the hub nodes, and prefetch/pin (or lock) the hub nodes in the main memory.
Consequently, we are proposing hub-node pinning strategies to pin recently used hub nodes
in the memory. To avoid wasting the memory resources and prevent high memory access
misses, we will predict and unpin recently unused pinned hub nodes with being aware of the
dynamic behavior of graph access events.
We will also propose a vertical integration solution to integrate the proposed crash-
deterministic SSD and validated SSD management system with dynamic graph applications,
which change nodes and edges through time. That approach of enhancing the performance
of SSDs by simplifying validation complexity and accounting for graph access behaviors will
ensure that graph data can be quickly and correctly recovered when a system encounters
sudden power failures or crashes. At the end of this three-year project, our GraphStor system
will integrate all of these research outputs, representing a highly e cient and reliable graph-
processing system.
105
